Multimodal radiological AI with LSO-Net++ for lumbar spine osteoporosis detection

Lumbar spine osteoporosis (LSO) is a chronic skeletal disorder that affects thousands of people worldwide, particularly older adults. Timely and accurate detection through radiological imaging is crucial; however, it poses medical challenges due to noise, subtle fractures, and limitations of available imaging modalities, particularly in low-dose X-ray images. Current state-of-the-art models face several limitations, including reliance on a single modality, computational inefficiency, insufficient interpretability, and limited generalization. This paper proposes a novel multimodal LSO-Net++ model for early osteoporosis detection across multiple modalities. First and foremost, we design and implement a novel noise-reduction module to enhance low-dose multimodal images without affecting edges or introducing blurriness. Following that, we used a state-of-the-art lightweight MobileNetV3 module to extract local features and a Swin Transformer for global contextual modelling. To overcome the generalization issue and leverage unlabelled images, we employed a SimCLR module. We have tested the performance of the proposed LSONet++ model across various multimodal, multicentre datasets. Experimental results demonstrate superior performance, robustness, and clinical application potential, with an F1-score of 0.94, sensitivity of 96.2%, and AUC of 0.98. Overall, LSO-Net++ represents a significant advancement in scalable and reliable AI-assisted osteoporosis detection using radiological imaging, particularly in challenging low-dose X-ray processing.